Relative depth estimation from dense matches using Graph Neural Networks

Sammanfattning

This thesis explores the application of Graph Neural Networks (GNNs) for relative depth estimation from dense matches between images. The main goals were to investigate the feasibility of using dense matches alone for this task, assess the performance improvement offered by GNNs over classical methods, and identify the most suitable features. The methodology involved utilizing the matches from Dense Kernelized Feature Matching (DKM) algorithm applied to the ScanNet-1500 and MegaDepth-1500 datasets. A multi-scale Graph Attention Network (MultiScaleGAT) model was developed and evaluated against Uniform-One and K-Nearest Neighbour baselines using Mean Squared Error (MSE) and Success Rates (SR) metrics. Experimental results demonstrate that dense correspondences provide sufficient information for relative depth estimation, and the MultiScaleGAT model significantly outperforms baseline methods on both indoor and outdoor datasets. A feature ablation study showed that the KNN ratio feature is crucial for cross-domain accuracy. The findings confirm the considerable advantage of using a learnable graph-based architecture for this task.